The scaling diagram has the same basic curve on the write-heavy test, but with fewer transactions occurring:
If you look carefully at the left side of this, you can see one of the issues warned about above and in the pgbench documentation: results at a very low database scale are lower than when the database is a bit larger. This is due to lock contention for the small branches table in particular, when running at higher client counts with a tiny scale.
The preceding results are averaged across all the client counts used. The number of clients also impacts results quite a bit:
Corresponding, the client results are also averaged ...